A set of observing system simulation experiments was performed.
This assessed the impact on global ocean biogeochemical reanalyses of assimilating chlorophyll from remotely sensed ocean colour and in situ observations of chlorophyll, nitrate, oxygen, and pH from a proposed array of Biogeochemical-Argo (BGC-Argo) floats.
Two potential BGC-Argo array distributions were tested: one for which biogeochemical sensors are placed on all current Argo floats and one for which biogeochemical sensors are placed on a quarter of current Argo floats.
Assimilating BGC-Argo data greatly improved model results throughout the water column.
This included surface partial pressure of carbon dioxide (
The works published in this journal are distributed under the Creative Commons Attribution 4.0 License. This license does not affect the Crown copyright work, which is re-usable under the Open Government Licence (OGL). The Creative Commons Attribution 4.0 License and the OGL are interoperable and do not conflict with, reduce or limit each other. © Crown copyright 2021
Throughout the ocean, physical and chemical processes interact with a teeming ecosystem to affect all life on Earth.
The upwelling of nutrient-rich waters fuels the growth of phytoplankton, which form the base of the marine food web and contribute half the planet's primary production
Understanding, monitoring, and predicting these processes is key to addressing some of the biggest challenges facing humanity.
Rising carbon dioxide (
Comprehensively monitoring all relevant processes in the global ocean, and their variability and trends, is not a trivial task.
For ocean biogeochemistry, the global observing system consists of various components which, while often sparse and disparate, have allowed fundamental insights.
More than 2 decades of routine satellite ocean colour data
Observation of ocean physics has been revolutionised by the advent of Argo
The Argo initiative is now being extended to biogeochemistry through the Biogeochemical-Argo (hereafter BGC-Argo) programme
The value of observations can be further enhanced by combining them with numerical models using data assimilation
The increasing availability of BGC-Argo data promises to change this, with great potential for improving reanalyses and forecasts
This paper describes the development of a scheme to assimilate profiles of Chl
The biogeochemistry OSSEs consider two potential scenarios: (1) a global BGC-Argo array equivalent to having biogeochemical sensors on one in four existing Argo floats, which is comparable to the planned target of 1000 floats, and (2) a global BGC-Argo array equivalent to having biogeochemical sensors on all existing Argo floats.
The aims were to assess the impact on reanalysis and forecasting systems that might be seen by assimilating multivariate BGC-Argo data, the influence of array size, and the value BGC-Argo would add to the existing ocean colour satellite constellation.
Assimilation of physics variables was not included due to the issues mentioned above, reflecting the way state-of-the-art biogeochemical reanalyses are run
This paper describes the updated model, newly developed assimilation scheme, and set-up of the OSSEs. Results are then presented showing the impact of assimilating the two potential BGC-Argo arrays, with and without ocean colour data. Finally, recommendations are made for the future development of observing and assimilation systems.
The reanalysis system is an upgraded version of that used in previous biogeochemical data assimilation studies at the Met Office
The physical ocean model used is the GO6 configuration
The biogeochemical ocean model used in this study is version 2 of the Model of Ecosystem Dynamics, nutrient Utilisation, Sequestration and Acidification (MEDUSA)
The data assimilation scheme used here is version 5 of NEMOVAR
When applied to physics data, NEMOVAR decomposes the full multivariate background error covariance matrix into an unbalanced and a balanced component for each variable.
The unbalanced component considers the uncorrelated component of each variable using univariate error covariances, while the balanced component considers correlations between variables.
The balanced component is derived using a set of linearised balance operators based on physical relationships
All increments are applied to the model over 1 d using incremental analysis updates (IAU)
NEMOVAR is used in this study to assimilate simulated ocean colour and BGC-Argo data, as described in the following sections.
NEMOVAR can be used for combined physical–biogeochemical assimilation
NEMOVAR was used here to assimilate total surface
For surface data, such as ocean colour, NEMOVAR can be applied in one of two ways.
The first, which is computationally most efficient and has been used in previous ocean colour assimilation studies
This gives a set of 3-D
For in situ profiles of biogeochemistry, as might be obtained from BGC-Argo, sets of 3-D increments were calculated for each assimilated variable, following the physics implementation of
In this study Chl
The Chl
Simulated BGC-Argo float trajectories for 2009 equivalent to having biogeochemical sensors on
Monthly mean surface
Absolute difference for December 2009 for surface
Annual zonal mean sections of
As detailed by a “nature run”, which is a realistic non-assimilative model simulation of the real world that provides a “truth” against which to validate the assimilative model; synthetic observations representing both existing routine observations and future observing networks, which are sampled from the nature run with appropriate errors prescribed; optionally, a free run, which provides an alternative model simulation of the nature run period; an assimilative run, which assimilates synthetic observations representing existing routine observations into the alternative model simulation; one or more additional versions of the assimilative run which also assimilate synthetic observations representing the future observing networks under consideration; and assessment of the impact on reanalysis or forecast skill of assimilating these observations by validating against the nature run.
One of the keys to obtaining informative results from an OSSE is to ensure that all sources of error are appropriately accounted for
In “identical twin experiments”, the nature and free runs differ only in their initial conditions.
This set-up was frequently used in early OSSEs, but as most sources of model error are neglected, the results were found to be overly optimistic, and the approach is no longer widely recommended In “fraternal twin experiments”, the same model is still used for both the nature and free run, but more aspects are modified.
These could include the initial conditions, lateral and surface boundary conditions, parameterisations, and resolution.
This takes much better account of model errors, and the approach is recommended over identical twin experiments In “full OSSEs”, significantly different models are used for the nature and free runs in order to make the two more independent.
The nature run is often run either at higher resolution than the assimilative model or with significantly different parameterisations
Due to the lack of availability of an appropriate alternative model for the nature run, it was decided within AtlantOS to take a fraternal twin approach for the biogeochemical OSSEs. This is sufficient to account for most sources of error, as long as any limitations of the approach are considered when drawing and acting upon conclusions.
The nature run in this study was run from 1 January 2008 to 31 December 2009 using the default parameterisations for the model versions used.
This is intended to be the best available non-assimilative model representation of the real world.
Validation of the general performance of the different system components can be found in the references given in Sect. 2, and validation of the nature run is presented in Sect. 4.1.
Atmospheric boundary conditions were taken from the ERA-Interim reanalysis
The free run was performed for the same period, including spin-up, but differed from the nature run in the following ways.
Atmospheric boundary conditions were taken from the JRA-55 reanalysis NEMO initial conditions were taken from an earlier date (1 January 1999) of the hindcast of MEDUSA initial conditions were taken from an earlier year (1218) of the UKESM1 ocean-only spin-up, with DIC and alkalinity taken from the end of the non-assimilative The NEMO parameter rn_efr, which affects near-inertial wave breaking and therefore vertical mixing The scheme used for advection of biogeochemical variables was changed from total variance dissipation (TVD) An alternative set of MEDUSA parameters was used, specifically parameter set 3 from Table 2 of
Together, these modifications generate approximations to the errors that exist in atmospheric fluxes and simulations of ocean physics and biogeochemistry. It is important to modify all of these, as errors in atmosphere and ocean physics have significant impacts on biogeochemical reanalyses and forecasts, and these errors must be accounted for if realistic conclusions are to be drawn from the OSSEs.
Synthetic ocean colour and BGC-Argo observations were generated from the nature run for each day of 2009.
Total Chl
In data assimilation, two components of observation error are typically considered: measurement error and representation error
Representation error arises from observations and models representing differing spatial and temporal scales and processes.
Since the nature and free runs were at the same resolution, this was accounted for in the same way as for the physics OSSEs in AtlantOS
For assimilating ocean colour data, the monthly varying background and observation error standard deviations from
For other variables, pre-existing error standard deviations were not available, so they were calculated for this study.
Observation error standard deviations were set to a climatological constant equal to the average global observation error specified.
These were fixed in time and specified as 0.638
Using these inputs, a set of assimilation experiments was performed in addition to the nature and free runs, as detailed in Table 1. The nature and free runs were run from 1 January 2008 to 31 December 2009, with the first year treated as spin-up. Each assimilation experiment was run from 1 January 2009 to 31 December 2009 using initial conditions from the end of the free run spin-up and assimilating the synthetic observations into the version of the model used for the free run.
Experiments performed.
Five assimilation experiments were run.
One just assimilated ocean colour.
Two assimilated ocean colour in combination with the
All the experiments, with unique identifiers for each, are detailed in Table 1.
The main metrics used for assessment are the absolute and percentage reduction in median absolute error (MAE), respectively defined as
Where
The results are presented in two subsections below. The first assesses the ability of NATURE to capture key ocean features and how differences between NATURE and FREE compare to errors in real-world reanalyses. The second assesses the assimilation runs and the potential impact of assimilating BGC-Argo and ocean colour data.
As stated in Sect. 3, OSSEs yield the most reliable conclusions when all sources of real-world error have been appropriately accounted for
Figure 2 shows surface fields of temperature, Chl
For both physical and biogeochemical variables, NATURE captured the broad global distribution, with generally appropriate magnitudes.
There were some discrepancies, such as an overestimation of Chl
FREE also broadly captured these features, as expected from state-of-the-art models
Absolute difference between FREE and real-world observation-based products
Time series of daily global RMSE for surface
Profiles of global
Surface
For temperature (Fig. 3a, b), the absolute difference between FREE and NATURE was very similar in pattern to that between FREE and the EN4 analysis but slightly lower in magnitude in some regions.
This suggests that the perturbations applied to the physics (different atmospheric fluxes, initial conditions, and vertical mixing) resulted in an error contribution to the biogeochemical model similar to, but slightly smaller than, that seen in state-of-the-art modelling systems.
For Chl
A similar comparison is required for the subsurface ocean, as BGC-Argo profiles were simulated for the upper 2000
A further consideration is that the growth of errors with time between FREE and NATURE should be comparable to that between FREE and real-world observations
While it is not ideal that the Chl
For each of the five assimilation runs, profiles of global
For Chl
The results for phytoplankton biomass were very similar to those for Chl
Surface
Hovmöller diagram of daily global
Hovmöller diagram of daily
Hovmöller diagram of daily
For
With the carbon cycle, DIC, alkalinity, and pH were all degraded in OC.
In ARGO_
Spatial maps of surface
Surface
The story for
Current state-of-the-art reanalyses typically assimilate ocean colour data
To investigate the impact of the BGC-Argo assimilation over the full year, a Hovmöller diagram
In Fig. 11, global
A set of observing system simulation experiments (OSSEs) was performed to explore the impact on global ocean biogeochemical reanalyses of assimilating currently available ocean colour data and assess the potential impact of assimilating BGC-Argo data.
Two different potential BGC-Argo array distributions were tested: one for which biogeochemical sensors are placed on all current Argo floats and one for which biogeochemical sensors are placed on a quarter of current Argo floats.
This latter approximately corresponds to the proposed BGC-Argo array of 1000 floats
Real-world experiments assimilating Chl
Adding assimilation of BGC-Argo profiles of Chl
For
There is much scope for improving data assimilation methodologies to better use existing satellite data and sparse in situ observations, which could bring at least as much benefit as expanding observing systems.
Multivariate balancing and better integration with physics data assimilation may help improve unassimilated variables.
More effective ways of spreading information from sparse data, such as cross-covariances based on empirical orthogonal functions or derived from an ensemble assimilation scheme, should also be considered.
Related to this, the correlation length scales used by the assimilation should be appropriately tuned for biogeochemical variables.
In this study, a single horizontal correlation length scale based on the first baroclinic Rossby radius was used, varying from a value of 25
A novel method for assimilating pH was introduced in this study, following the method for assimilating
From the point of view of ocean data assimilation, BGC-Argo will bring significant advances in reanalysis and forecasting skill, and it is recommended to proceed with its development as a priority.
The proposed array of 1000 floats will be enough to deliver clear improvements, and a larger array would be likely to bring further benefits.
Ocean colour and BGC-Argo provide complementary information, so maintaining and developing the existing ocean colour satellite constellation should also be a priority.
Technologies such as gliders may also bring additional benefits, for instance for
The nature of the 4-D data generated in running the model experiments requires a large tape storage facility. These data are in excess of 100 terabytes (TB). However, the data can be made available upon request from the author.
The author declares that there is no conflict of interest.
This article is part of the special issue “Biogeochemistry in the BGC-Argo era: from process studies to ecosystem forecasts (BG/OS inter-journal SI)”. It is not associated with a conference.
The author would like to thank Susan Kay and Matt Martin for useful discussions and comments on the draft paper, as well as the two anonymous reviewers for their helpful comments in the
This study received funding from the European Union's Horizon 2020 Research and Innovation programme under grant agreement 633211 (AtlantOS).
This paper was edited by Katja Fennel and reviewed by two anonymous referees.